Uncertainty Estimates for Optical Flow with Multi-Hypotheses Networks
Recent work has shown that optical flow estimation can be formulated as an end-to-end supervised learning problem, which yields estimates with a superior accuracy-runtime tradeoff compared to alternative methodology. In this paper, we make the network estimate its local uncertainty about the correctness of its prediction, which is vital information when building decisions on top of the estimated optical flow. For the first time we compare several strategies and techniques to estimate uncertainty in a large-scale computer vision task like optical flow estimation. Moreover, we introduce a new network architecture that enforces complementary hypotheses and provides uncertainty estimates efficiently within a single forward pass without the need for sampling or ensembles. We demonstrate high-quality uncertainty estimates that clearly improve over previous confidence measures on optical flow and allow for interactive frame rates.
READ FULL TEXT